# A First Look at On-device Models in iOS Apps

**Authors:** Han Hu, Yujin Huang, Qiuyuan Chen, Terry Tue Zhuo, Chunyang Chen  
**arXiv:** [2307.12328](https://arxiv.org/abs/2307.12328)  
**Date:** 2023-07-23  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2307.12328v1/

## Key points

- Researchers explore factors influencing developers' choices of on-device models for iOS and Android apps
- Study involves analyzing 62 pairs of iOS-Android apps out of a total of 312
- Identify 226 on-device models with specific functionalities in selected iOS apps
- Pipeline includes identifying functionalities and usage scenarios in iOS apps, then locating similar scenarios in Android counterparts using code analysis techniques
- Semantic information like file names, method names, and user interface elements used to infer potential usage scenarios
- Findings validated by running apps to verify inferred scenarios align with model functionalities
- Three alternatives for on-device models: utilizing other on-device models, invoking on-cloud models, or implementing custom obfuscation techniques
- Developers may use cloud-based services by pre-configuring SDK libraries or resorting to customized DL models when no suitable options are available

## Summary

In their study published in ACM Trans. Softw. Eng. Methodol., researchers Han Hu, Yujin Huang, Qiuyuan Chen, Terry Yue zhuo, and Chunyang Chen explore the factors influencing developers' choices of on-device models for iOS and Android apps. They establish a comprehensive pipeline to analyze how developers replace and share on-device models between the two platforms by examining 62 pairs of iOS-Android apps out of a total of 312. Through manual analysis, they identify 226 on-device models with specific functionalities in the selected iOS apps. The study pipeline involves identifying the functionalities and usage scenarios of on-device models in iOS apps and then attempting to locate similar scenarios in Android counterparts using code analysis techniques. Semantic information such as file names, method names, and user interface elements are also analyzed to infer potential usage scenarios when direct code comparison is not feasible. The researchers validate their findings by running the apps to verify if inferred scenarios align with model functionalities. The study reveals three alternatives for on-device models: utilizing other on-device models for the same task, invoking on-cloud models, or implementing custom obfuscation techniques. Developers may choose to use cloud-based services by pre-configuring SDK libraries in Android apps or resorting to customized DL models when no suitable options are available. Overall, this research sheds light on the complexities involved in selecting on-device models for iOS and Android apps and provides insights into developers' decision-making processes when it comes to leveraging deep learning frameworks for mobile applications.